Stacking Machine Learning Models for Predicting Photophysical Properties of Iridium Complexes

Iridium complexes have played a key role in organic emitting diodes (OLEDs) devices as emitter for its high efficiency, appropriate solubility, and availability for emission color tuning. Although several progress has been made by using machine learning (ML) in designing new materials for OLEDs, there is still much room for enhancing the capability, transferability, and usability of machine learning. In the present work, we construct stacking ML models for predicting emission properties of iridium complexes, and multi-task for absorption wavelength prediction. Compared with the existing ML models, we propose an exhaustive feature generation process for iridium complexes, in which the original features of ligands of each complex are combined together according to the order of electronic structure properties. Then we build a series of meta learners in order to improve the accuracy in the emission task and to transfer ML models for predicting other photophysical properties such as the absorption wavelength. The influence of different feature selection methods and different machine learning algorithms on the performance of ML prediction is discussed in detail. We suggest concatenating the Morgan fingerprints of ligands according to a sorted order of electronic structure properties of isolated ligands as a simple and robust way to generate the input feature of complexes. The electronic structure calculations on isolated ligands can be replaced by ML, which further improves computational efficiency of the whole procedure. Promotion from base to meta learners and the advantage of gradient boosting algorithms can be observed in most tasks.